- Time
- Post link
So what's the latest ETA for this?
Sadly my projects are lost due to an HDD crash… 😦 | [Fundamental Collection] thread | blog.spoRv.com | fan preservation forum: fanres.com
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *) — Page 29
- Time
- Post link
We are all chomping at the bit! That's a good sign that you made a great breakthrough! :)
- Time
- Post link
stretch009 said:
So what's the latest ETA for this?
If he's just starting to render now, it will be at least a few days if it's going at as slow a pace as his older versions. That's assuming that a problem doesn't come up. I think it's probably been a while since the rendering process has completed the whole film?
- Time
- Post link
thorr said:
We are all chomping at the bit!
I wouldn't make horse references with Stinky-Dinkins lurking, if I were you.
- Time
- Post link
lol :-)
- Time
- (Edited)
- Post link
The eta is as soon as possible. The rendering takes about two weeks, which would be fine, if not for the fact that I had to start over with each iteration. ;-)
I do think this may be a good time to get a version out there, such that you can judge it's merits, which can be used as input for another version in the future.
- Time
- Post link
Great job DrDre!
Now it's time for us to fight about the 1080p version! (^^,)
Waiting for some "usual" frames, but at 1920x816...
- Time
- Post link
It probably would be best to get a full version out soon like you said. As there may be some unforeseen problems later in the film. Having multiple people scrutinize the film will allow for a comprehensive analysis of it.
I think there is merit to a '1080P' version! '720P' is just too small of a resolution to be very effective today. But considering the source, it does really well!
- Time
- Post link
Woah, those comparisons with the Spline 64 resize are more than incredible!
The anti-aliasing really makes it much easier to the eyes.
I have always been a defender of higher resolution, so I would opt in for 1080p too ;)
Darth Id on ‘Why “Ben”?’:
And while we’re at it, we need to figure out why they kept calling Mark Hamill’s character “Luke Skywalker,” since it’s my subjective opinion that his name is actually Schnarzle Shnuzzle. It just doesn’t make sense!
Damn you George Lucas for never explaining why they all keep calling Schnarzle “Luke”!
Damn You!!!
- Time
- Post link
I was wondering how chroma is being handled. A few small lights in your comparisons seem less colorful than Spline64, for instance the red lines on the wall in frame 4691. I would think you could just use the raw super resolution for chroma without artifacts being noticeable, but if not spline64 would be good enough
- Time
- Post link
I think that's more to do with deblurring. The red lights, but also the black lines are pretty blurry in the Spline64.
- Time
- Post link
I kinda thought that too, but was too lazy to do a real test.
http://screenshotcomparison.com/comparison/128400
look at the orange light to the left of 3po's head.
- Time
- Post link
http://screenshotcomparison.com/comparison/128263 The red light in this one has a more noticeable change. Doesn't bother me at all, but I am beginning to wonder how the lightsabers are going to look.
- Time
- (Edited)
- Post link
One of the problems is that many of the scripts and filters in Avisynth work in different color spaces, so there are a number of color space conversions. I've reduced the number of conversions to the least possible, but there are still a few. The worst of these is the YV12 conversion, which seems to be the main reason for the dimmed colors. I'll merge the chroma channel of SRV10 with the Spline64.
Edit:
I actually found a different way to correct the colors. Here's what frame 4961 and 7218 look like with the correct colors compared to the Spline64:
- Time
- Post link
To ease the fears about those lightsabers, here's an example:
- Time
- Post link
Here's one of those notoriously grainy shots on Tatooine and how it will be handled by SRV10:
- Time
- Post link
Great! Are you incorporating the new color fix into the latest video you will be providing?
- Time
- Post link
Yep, the video sample should be online in a few days.
- Time
- Post link
Yep, the minor color glitch looks fixed. As long as its not at the cost of any v10 detail.
- Time
- (Edited)
- Post link
No, the way I implemented the solution doesn't affect the detail. I substract a simple upscale with the same color conversions as the SRV10, and then add the simple upscale without the color conversions. This should by definition restore the color, without affecting anything else.
- Time
- (Edited)
- Post link
Did you tweak any other settings or add any kind of compression? because I'm seeing subtle luma differences in frame 7218 comparing your new sample with srv10 in photoshop. Though looking just at the chroma in the rgb channels, the new sample seems more detailed than both srv10 and spline64. I really hope you post your script in the end, sounds full of clever solutions.
- Time
- Post link
I applied very light sharpening to counter the blurring of edges by the antialiasing, and of course the color correction.
I will be posting the script, once the trilogy is finished. :-)
- Time
- (Edited)
- Post link
DrDre said:
Thanks for the compliments! Glad to hear you enjoy the work and screenshot comparisons. I also often use mobile devices for the screenshot comparisons, because the zooming on mobile devices makes it easier to look at the details, than to put your face against your computer screen.
As with most techniques, there are many ways to implement the methodology. However, as a rough guide, a super resolution algorithm defines a reference frame, and subsequently estimates how objects, often defined as image patches, can be mapped onto the reference frame. The image patches are then averaged using some form of weighted averaging, where the weights are related to the similarity to the reference frame. The weighting reduces the chance of wrong detail placement, the socalled registration error. After the averaging the resulting frame is deblurred leading to the final result.
DrDre said:
I will try to create a video sample as soon as possible.
1920-wide upscales are possible, only it won't add much in terms of detail, so an HD-TV's internal upscaler will probably approach the same quality.
Thank you for you detailed replies! One thought that I had is that multiplying the Standard Definition picture area by 4 would closely approximate "Full HDTV" 1920x1080 dimensions (though non-anamorphic black letterbox bars will not be contributing at all to vertical resolution). Would this not make the enlargement process computationally simpler than upscaling ~480 pixels across to 1280 across? A factor of exactly 4 (4 x 480 = 1920) seems as though it would be simpler (and perhaps visually "cleaner") for the computer to map pixels, rather than a factor of ~2-2/3. The current method, though producing amazing results with a smaller output file, seems like there would be some heavier computation and possibly a SR upscale slightly less than its full potential, due to the ~2.66 x <480 computation.
EDIT: Clearly I had a major fail with my horizontal resolution spec. 480 is a vertical resolution spec.
If your crop is water, what, exactly, would you dust your crops with?
- Time
- (Edited)
- Post link
The actual video being upscaled is 712x274. The SR algorithm upscales the video exactly two times to 1424x548. At this point I downscale the video the 1280x544 with Spline64, but I might upscale to 1920x816 for the final version.
- Time
- Post link
DrDre said:
The actual video being upscaled is 712x274. The SR algorithm upscales the video exactly two times to 1424x548. At this point I downscale the video the 1280x544 with Spline64, but I might upscale to 1920x816 for the final version.
How about 1920x1080 for those who want to burn it directly to a BD?